ARTIFICIAL INTELLIGENCE

handsscandalousAI and Robotics

Jul 18, 2012 (5 years and 3 months ago)

283 views

ARTIFICIAL INTELLIGENCE 17
Rete: A Fast Algorithm for the
Many Pattern/Many Object
Pattern Match Problem*
Charl es L. Forgy
Department of Computer Science, Carnegie-Mellon University,
Pittsburgh, PA 15213, U.S.A.
Recommended by Harry Barrow
ABSTRACT
The Rete Match Algorithm is an efficient method for comparing a large collection of patterns to a large
collection of objects. It finds all the objects that match each pattern. The algorithm was developed for use in
production system interpreters, and it has been used for systems containing from a few hundred to more
than a thousand patterns and objects. This article presents the algorithm in detail. It explains the basic
concepts of the algorithm, it describes pattern and object representations that are appropriate for the
algorithm, and it describes the operations performed by the pattern matcher.
1. Introduction
In many pattern/many object pattern matching, a collection of patterns is
compared to a collection of objects, and all the matches are determined. That
is, the pattern matcher finds every object that matches each pattern. This kind
of pattern matching is used extensively in Artificial Intelligence programs
today. For instance, it is a basic component of production system interpreters.
The interpreters use it to determine ~vhich productions have satisfied condition
parts. Unfortunately, it can be slow when large numbers of patterns or objects
are involved. Some systems have been observed to spend more than nine-
tenths of their total run time performing this kind of pattern matching [5]. This
*This research was sponsored by the Defense Advanced Research Projects Agency (DOD),
ARPA Order No. 3597, monitored by the Air Force Avionics Laboratory under Contract
F33615-78-C-1551.
The views and conclusions contained in this document are those of the author and should not be
interpreted as representing the ottieial policies, either expressed or implied, of the Defense
Advanced Research Projects Agency or the US Government.
Artificial Intelligence 19 (1982) 17-37
IX104-3702/82/~/$02.75 O 1982 North-Holland
18 C.L. FORGY
article describes an algorithm that was designed to make many pattern/many
object pattern matching less expensive. The algorithm was developed for use in
production system interpreters, but since it should be useful for other lan-
guages and systems as well, it is presented in detail.
This article attends to two complementary aspects of efficiency: (1) designing
an algorithm for the task and (2) implementing the algorithm on the computer.
The rest of Section 1 provides some background information. Section 2
presents the basic concepts of the algorithm. Section 3 explains how the objects
and patterns should be represented to allow the most efficient implementations.
Section 4 describes in detail a very fast implementation of the algorithm.
Finally, Section 5 presents some of the results of the analyses of the algorithm.
1.1. oPs5
The methods described in this article were developed for production system
interpreters, and they will be illustrated with examples drawn from production
systems. This section provides a brief introduction to the language used in the
examples, oPs5. For a more complete description of oPS5, see [6].
A production system program consists of an unordered collection of If-Then
statements called productions. The data operated on by the productions is held
in a global data base called working memory. By convention, the If part of a
production is called its LHS (left-hand side), and its Then part is called its RHS
(right-hand side). The interpreter executes a production system by performing
the following operations.
(1) Match. Evaluate the LHSs of the productions to determine which are
satisfied given the current contents of working memory.
(2) Conflict resolution. Select one production with a satisfied LHS; if no
productions have satisfied LHSs, halt the interpreter.
(3) Act. Perform the actions in the RHS of the selected production.
(4) Goto 1.
oPss working memories typically contain several hundred objects, and each
object typically has between ten and one hundred associated attribute-value
pairs. An object together with its attribute-value pairs is called a working
memory element. The following is a typical, though very small, oPs5 working
memory element; it indicates that the object of class Expression which is
named Exprl7 has 2 as its first argument, '*' as its operator, and X as its second
argument.
(Expression 1' Name Exprl7 1' Argl 2 1' Op * ~' Arg2 X)
The 1' is the oPs5 operator that distinguishes attributes from values.
The LHS of a production consists of a sequence of patterns; that is, a
sequence of partial descriptions of working memory elements. When a pattern
P describes an element E, P is said to match E. In some productions, some of
THE RETE MATCH ALGORITHM 19
the patterns are preceded by the negation symbol, -. An LHS is satisfied when
(1) Every pattern that is not preceded by-mat ches a working memory
element, and
(2) No pattern that is preceded by-mat ches a working memory element.
The simplest patterns contain only constant symbols and numbers. A pattern
containing only constants matches a working memory element if every constant
in the pattern occurs in the corresponding position in the working memory
element. (Since patterns are partial descriptions, it is not necessary for every
constant in the working memory element to occur in the pattern.) Thus the
pattern
(Expression 1' Op * 1' Arg2 0)
would match the element
(Expression 1' Name Expr86 1' Argl X 1' Op * 1` Arg2 0)
Many non-constant symbols are available in oPS5 for definfng patterns, but the
two most important are variables and predicates. A variable is a symbol that
begins with the character '(' and ends with the character ')'--for example (X). A
variable in a pattern will match any value in a working memory element, but if a
variable occurs more than once in a production's LHS, all occurrences must match
the same value. Thus the pattern
(Expression T Argl (VAL) T Arg2 (VAL))
would match either of the following
(Expression 1' Name Expr9 1`Argl Expr23 1'Op * TArg2
Expr23)
(Expression 1` Name Expr5 1`Argl 0 1`Op- l'Arg2 0)
but it would not match
(Expression 1' Name Expr8 1' Argl 0 1' Op * 1' Arg2 Expr23)
The predicates in oPs5 include = (equal), <> (not equal), < (less than), >
(greater than), <= (less than or equal), and >= (greater than or equal). A
predicate is placed between an attribute and a value to indicate that the value
matched must be related in that way to the value in the pattern. For instance,
(Expression 1' Op <>*)
will match any expression whose operand is not *. Predicates can be used with
variables as well as with constant values. For example, the following pattern
(Expression 1' Argl (LEFT) 1' Arg2 <> (LEFT))
will match any expression in which the first argument differs from the second
argument.
20 C.L. FORGY
The RHS of a production consists of an unconditional sequence of actions.
The only actions that need to be described here are the ones that change
working memory. MAKE builds a new element and adds it to working
memory. The argument to MAKE is a pattern like the patterns in LHSs. For
example,
(MAKE Expression ~' Name Exprl T Argl 1)
will build an expression whose name is Exprl, whose first argument is I, and
whose other attributes all have the value NIL (the default value in oess).
MODIFY changes one or more values of an existing element. This action takes
as arguments a pattern designator and a list of attribute-value pairs. The
following action, for example
(MODIFY 2 ~' Op NIL T Arg2 NIL)
would take the expression matching the second pattern and change its operator
and second argument to NIL. The action REMOVE deletes elements from
working memory. It takes pattern designators as arguments. For example
( REMOVE 1 2 3)
would delete the elements matching the first three patterns in a production.
An or's5 production consists of (1) the symbol P, (2) the name of the
production, (3) the LHS, (4) the symbol - - >, and (5) the RHS, with everything
enclosed in parentheses. The following is a typical production.
(P Time 0x
(Goal 1' Type Simplify 1' Object (X))
(Expression I' Name (X) 1' Argl 0 1' Op *)
(MODIFY 2 I'OP NIL ~'Arg2 NIL))
1.2. Work on production system efficiency
Since execution speed has always been a major issue for production systems,
several researchers have worked on the problem of efficiency. The most
common approach has been to combine a process called indexing with direct
interpretation of the LHSs. In the simplest form of indexing, the interpreter
begins the match process by extracting one or more features from each working
memory element, and uses those features to hash into the collection of
productions. This produces a set of productions that might have satisfied LHSs.
The interpreter examines each LHS in this set individually to determine
whether it is in fact satisfied. A more efficient form of indexing adds memory to
the process. A typical scheme involves storing a count with each pattern. The
counts are all zero when execution of the system begins. When an element
enters working memory, the indexing function is executed with the new
THE RETE MATCH ALGORrrHM 21
element as its only input, and all the patterns that are reached have their
counts increased by one. When an element leaves working memory, the index
is again executed, and the patterns that are reached have their counts
decreased by one. The interpreter performs the direct interpretation step only
on those LHSs that have non-zero counts for all their patterns. Interpreters
using this scheme--in some cases combined with other efficiency measures--
have been described by McCracken [8], McDermott, Newell, and Moore [9],
and Rychener [10].
The algorithm that will be presented here, the Rete Match Algorithm, can be
described as an indexing scheme that does not require the interpretive step.
The indexing function is represented as a network of simple feature recog-
nizers. This representation is related to the graph representations for so-called
structured patterns. (See for example [2] and [7]). The Rete algorithm was first
described in 1974 [3]. A 1977 paper [4] described some rather complex
interpreters for the networks of feature recognizers, including parallel inter-
preters and interpreters which delayed evaluation of patterns as long as
possible. (Delaying evaluation is useful because it makes it less likely that
patterns will be evaluated unnecessarily.) A 1979 paper [5] discussed simple but
very fast interpreters for the networks. This article is based in large part on the
1979 paper.
2. The Rete Match Algorithm---Basic Concepts
In a production system interpreter, the output of the match process and the
input to conflict resolution is a set called the conflict set. The conflict set is a
collection of ordered pairs of the form
(Production, List of elements matched by its LHS)
The ordered pairs are called instantiations. The Rete Match Algorithm is an
algorithm for computing the conflict set. That is, it is an algorithm to compare a
set of LHSs to a set of elements fn order to discover all the instantiations. The
algorithm can efficiently process large sets because it does not iterate over the
sets.
2.1. How to avoid iterating over working memory
A pattern matcher can avoid iterating over the elements in working memory by
storing information between cycles. The step that can require iteration is
determining whether a given pattern matches any of the working memory
elements. The simplest interpreters determine this by comparing the pattern to
the elements one by one. The iteration can be avoided by storing, with each
pattern, a list of the elements that it matches. The lists are updated when
working memory changes. When an element enters working memory, the
interpreter finds all the patterns that match it and adds it to their lists. When an
22 C.L. FORGY
element leaves working memory, the interpreter again finds all the patterns
that match it and deletes it from their lists.
Since pattern matchers using the Rete algorithm save this kind of in-
formation, they never have to examine working memory. The pattern matcher
can be viewed as a black box with one input and one output.
(Changes to Working Memory)
1
Black Box
1
(Changes to the Conflict Set)
The box receives information about the changes that are made to working
memory, and it determines the changes that must be made in the conflict set to
keep it consistent. For example, the black box might be told that the element
(Goal T Type Simplify 1' Object Exprl9)
has been added to working memory, and it might respond that production
TimexN has just become instantiated.
2.1.1. Tokens
The descriptions of working memory changes that are passed into the black
box are called tokens. A token is an ordered pair of a tag and a list of data
elements. In the simplest implementations of the Rete Match Algorithm, only
two tags are needed, + and -. The tag + indicates that something has been
added to working memory. The tag - indicates that something has been
deleted from working memory. When an element is modified, two tokens are
sent to the black box; one token indicates that the old form of the element has
been deleted from working memory, and the other that the new form of the
element has been added. For example, if
(Expression 1' Name Expr41 1' Argl Y
was changed to
(Expression 1' Name Expr41 I' Argl 2
the following two tokens would be processed.
(-(Expression
(+(Expression
1' Op + 1' Arg2 Y)
TOp * I' Arg2 Y)
1' Name Expr41 T Argl Y 1" Op + 1' Arg2 Y)>
TNameExpr41 l'Ar gl 2 I'Op* l'Arg2Y)>
2.2. How to avoid iterating over production memory
The Rete algorithm avoids iterating over the set of productions by using a
tree-structured sorting network or index for the productions. The network,
THE RETE MATCH ALGORITHM 23
which is compiled from the patterns, is the principal component of the black
box. The following sections explain how patterns are compiled into networks
and how the networks perform the functions of the black box.
2.2.1. Compiling the patterns
When a pattern matcher processes a working memory element, it tests many
features of the element. The features can be divided into two classes. The first
class, which could be called the intra-element features, are the ones that
involve only one working memory element. For an example of these features,
consider the following pattern.
(Expression 1' Name (N) l'Argl 0 l'Op + l'Arg2 (X))
When the pattern matcher processes this pattern, it tries to find working
memory elements having the following intra-element features.
-The class of the element must be Expression.
-The value of the Argl attribute must be the number 0.
- The value of the Op attribute must be the atom +.
The other class of features, the inter-element features, results from having a
variable occur in more than one pattern. Consider Plus0x's LHS.
(P Plus0x
(Goal 1' Type Simplify 1' Object (N))
(Expression 1' Name (N) l'Argl 0 l'Op +
_- ~ . . .)
1' Arg2 (X))
The intra-element features for the second pattern are listed above. A similar
list can be constructed for the first pattern. But in addition to those two lists,
the following inter-element feature is necessary because the variable (N) occurs
twice.
-The value of the Object attribute of the goal must be equal to the value of
the Name attribute of the expression.
The pattern compiler builds a network by linking together nodes which test
elements for these features. When the compiler processes an LHS, it begins
with the intra-element features. It d~termines the intra-element features that
each pattern requires and builds a linear sequence of nodes for the pattern.
Each node tests for the presence of one feature. After the compiler finishes
with the intra-element features, it builds nodes to test for the inter-element
features. Each of the nodes has two inputs so that it can join two paths in the
network into one. The first of the two-input nodes joins t he linear sequences
for the first two patterns, the second two-input nodes joins the output of the
first with the sequence for the third pattern, and so on. The two-input nodes
test every inter-element feature that applies to the elements they process.
Finally, after the two-input nodes, the compiler builds a special terminal node
to represent the production. This node is attached to the last of the two-input
24 C.L. FORGY
nodes. Fig. 1 shows the network for Plus0x and the similar production Time0x.
Note that when two LHSs require identical nodes, the compiler shares parts of
the network rather than building duplicate nodes.
2.2.2. Processing in the network
The root node of the network (at the top in Fig. 1) is the input to the black box.
This node receives the tokens that are sent to the black box and passes copies
of the tokens to all its successors. The successors of the top node, the nodes to
perform the intra-element tests, have one input and one or more outputs. Each
node tests one feature and sends the tokens that pass the test to its successors.
The two-input nodes compare tokens from different paths and join them into
bigger tokens if they satisfy the inter-element constraints of the LHS. Because
of the tests performed by the other nodes, a terminal node will receive only
tokens that instantiate the LHS. The terminal node sends out of the black box
the information that the conflict set must be changed.
For an example of the operation of the nodes, consider what happens in the
network in Fig. 1 when the following two elements are put into an empty
working memory.
(Goal 1' Type Simplify 1' Object Exprl7)
(Expression 1' Name Expr17 1' Argl 0 1' Op * 1' Arg2 X)
First the token
(+(Goal 1' Type Simplify 1' Object Exprl7))
is created and sent to the root of the network. This node sends the token to its
successors. One of the successors (on the right in Fig. 1) tests it and rejects it
because its class is not Expression. This node does not pass the token to its
successor. The other successor of the top node accepts the token (because its
class is Goal) and so sends it to its successor. That node also accepts the token
(since its type is Simplify), and it sends the token to its successors, the
two-input nodes. Since no other tokens have arrived at the two-input nodes,
they can perform no tests; they must just store the token and wait.
When the token
(+(Expression 1' Name Exprl7 1' Argl 0 1` Op * 1' Arg2 X))
is processed, it is tested by the one-input nodes and passed down to the right
input of Time0x's two-input node. This node compares the new token to the
earlier one, and finding that they allow the variable to be bound consistently, it
creates and sends out the token
(+(Goal 1' Type Simplify 1' Object Expr17)
(Expression 1' Name Expr17 1' Argl 0 1' Op * 1' Arg2 X))
THE RETE MATCH ALGORITHM 25
(P Plus0x
(Goal 1" Type Simplify
(Expression T Name (N)
__> . . .)
(P Time0x
(Goal 1" Type Simplify
(Expression 1" Name (N)
__:> . . .)
1" Object (N))
1"Argl 0 1"Op+
1" Object (N))
1"Argl0 1"Op*
1" Arg2 (X))
1" Arg2 (X))
Distribute the tokens.
j l
Is the element class Goal?
Is the value of the
Type Simplify? ~
Is the value of the
Op+? [
Join the elements in which
the value of the Object
attribute from the left
is equal to
the value of the Name
attribute from the right.
Report that production
Plus0x is satisfied.
Is the element class Expression?
Is the value of the Argl 0?
Is the value of the
l
Join the elements in which
the value of the Object
attribute from the left
is equal to
the value of the Name
attribute from the right.
Report that production
Time0x is satisfied.
FIG. 1. The network for Plus0x and Time0x.
When its successor, the terminal node for Time0x, receives this token, it adds
the instantiation of Time0x to the conflict set.
2.2.3. Saving information in the network
As explained above, the black box must maintain state information because it
must know what is in working memory. In simple Rete networks all such state
is stored by the two-input nodes. Each two-input node contains two lists called
its left and right memories. The left memory holds copies of the tokens that
arrived at its left input, and the right memory holds copies of the tokens that
arrived at its right input. The tokens are stored as long as they are useful. The
next section explains how the nodes determine when the tokens are no longer
useful.
26 C.L. FORGY
2.2.4. Using the tags
The tag in a token indicates how the state information is to be changed when
the token is processed. The + and - tokens are processed identically except:
- The terminal nodes use the tags to det ermi ne whether to add an instantiation to
the conflict set or to remove an existing instantiation. When a + token is
processed, an instantiation is added; when a - token is processed, an
instantiation is removed.
- The two-input nodes use the tags to det ermi ne how to modify their internal
memories. When a + token is processed, it is stored in the internal memory;
when a - token is processed, a token with an identical data part is deleted.
- The two-input nodes use the tags to det ermi ne the appropriate tags for the
tokens they build. When a new output is created, it is given the tag of the token
that just arrived at the two-input node.
2.3. Completing the set of node types
The network in Fig. 1 contained four kinds of nodes: the root node, the
terminal nodes, the one-input nodes, and the two-input nodes. Certainly one
could define many more kinds of nodes, but only a few more are necessary to
have a compl et e and useful set. In fact, only two more kinds of nodes are
necessary to interpret oPs5.
A second kind of two-input node is needed for negated patterns (that is,
patterns preceded by - ). The new two-input node stores a count with each
token in its left memory. The count indicates the number of tokens in the right
memory that allow consistent variable bindings. The tokens in its right memory
contain the elements that match the negated pattern---or, more precisely, the
tokens contain the elements that have the intra-element features that the
negated pattern requires. The node allows the tokens with a count of zero to
pass.
The last node type that needs to be defined is a variant of the one-input
nodes described earlier. Those nodes tested Working memory elements for
constant features (testing, for example, whether a value was equal to a given
atomic symbol). The new one-input nodes compare two values from a working
memory element. These nodes are used to process patterns that contain two or
more occurrences of a variable. The following, for example, would require one
of these nodes because (X) occurs twice.
(Expression 1' Ar gl (X) 1' Op + 1' Arg2 (X))
3. Representing the Network and the Tokens
This section describes representations for t okens and nodes,,that allow very fast
interpreters to be written.
THE RETE MATCH ALGORITHM 27
3.1. Working memory elements
The representation chosen for the working memory elements should have two
properties.
- The representation should make it easy to extract values from elements because
every test involves extracting one or more values.
- The representation should make it easy to perform the tests once the values are
available.
To make extracting the values easy, each element should be stored in a
contiguous block in memory, and each attribute should have a designated index
in the block. For example, if elements of class Ck had seventeen attributes, A1
through A17, they should be stored as blocks of eighteen values. The first value
would be the class name (Ck). The second value would be the value of attribute
A1. The third would be the value of attribute A2, and so on. The particular
assignment of indices to attributes is unimportant; it is important only that each
attribute have a fixed index, and that the indices be assigned at compile time.
This allows the compiler to build the indices into the nodes. Thus instead of a
node like the following:
Is the value of the Status attribute Pending?
the compiler could build the node
Is the value at location 8 Pending?
With this representation, each value can be accessed in one memory reference,
regardless of the number of attributes possessed by an element.
To make the tests inexpensive, the representation should have explicit type
bits. One obvious way to represent a value is to use one word for the type and
one or more words for the value proper. But more space-efficient represen-
tations are also possible. For example, consider a production system language
that supports three data types, integers, floating point numbers, and atoms. A
representation like the following might be used: One word would be allocated
to each value. For integers and atoms, the low order sixteen (say) bits would
hold the datum and the seventeenth bit would be a type bit. For floating point
numbers, the entire word would be used to store a normalized floating point
number. A floating point number would be recognized by having at least one
non-zero in the high order bits.
3.2. The network
This section explains how to represent nodes in a form similar to von Neumann
machine instructions. This representation was chosen because it allows the
network interpreter to be organized like the interpreters for conventional von
Neumann architectures.
28 C.L. FORGY
3.2.1 An assembly language notation
To make it easier to discuss the representation for the nodes, an assembly
language notation is used below A one-input node like
Is the value of locating 8 Pending?
becomes
TEQA 8, Pending
The T, which stands for test, indicates that this is a one-input node The EQ
indicates that it is a test for equality (It is also necessary to have NE for not
equals, LT for less than, etc.). The A indicates the node tests data of type
atom. (There is also a type N for integer values, a type F for floating point, and
a type S for comparing two values in the same working memory element).
Two-input nodes are indicated by lines like the following
L001 AND (2)= (1)
L001 is a label. AND indicates that this is a two-input node for non-negated
patterns. The sequence (2)= (1) indicates that the node compares the second
value of elements from the left and the first value of elements from the right;
the -- indicates that it performs a test for equality. The terminal nodes contain
the type TERM and the name of the production For example
TERM Plus0x
As will be explained below, the ROOT node is not needed in this represen-
tation
322 Linearizing the network
To make the nodes like the instructions for avon Neumann machine, it is
necessary to eliminate the explicit links between nodes. Many of the explicit
links can be eliminated simply by linearizing the network, placing a node and
its successor in contiguous memory locations However, since some nodes have
more than one successor, and others (the two-input nodes) have more than one
predecessor, linearizing is not sufficient in itself: two new node types must be
defined to replace some of the links. The first of the new nodes, the FORK, is
used to indicate that a node has more than one successor. The FORK node
contains the address of one of the successors. The other successor is placed
immediately after the FORK For example, the FORK in the following
indicates that the node L003 has two successors
L003 TEQA 0, Expression
FORK L004
TEQA 3, +
L004 TEQA 3, *
THE RETE MATCH ALGORITHM 29
The other new node type, the MERGE, is used where the network has to grow
back t oget her--t hat is, before two-input nodes. The two-input node is placed
after one of its predecessors (say its left predecessor) and the MERGE is
placed after the other. The MERGE, which contains the address of the
two-input node, functions much like an unconditional jump. Fig. 2 shows the
effect of the linearization process; it contains the productions from Fig. 1 and
the linearized network for their LHSs.
3.2.3. Representing the nodes in memory
This section shows how the nodes could be represented on a computer which
has a thirty-two bit word length. The thirty-two bit word length was chosen
because it is typical of today's computers; the precise word length is not
critical, however. Since the network can be rooted at a FORK (see the example
in Fig. 2) it is not necessary to have an explicit root node for the network.
Hence only seven classes of nodes are needed; FORKs, MERGEs, the two
kinds of one-input nodes, the two kinds of two-input nodes, and the terminal
nodes.
FORKs and MERGEs could be represented as single words. Six bits could
be used for a type field (that is, a field to indicate what the word represents)
and the remaining twenty-six bits could be used for the address of the node
pointed to. FORKs and MERGEs would thus be represented:
I I I
I TYPE [ ADDRESS I
I I I
I_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1_1
(6 bi t s) (26 bi t s)
Both kinds of one-input nodes could be represented as single words that are
divided into three fields. The first field would hold the type of the node. The
second field would hold the index of the value to test. The third field would
hold either a constant or a second index. The bits in a word could be allocated
as follows.
I I I I
I TYPE I INDEX I CONSTANT or INDEX I
i I I I
l _l _l _I_I_I_l _l _l _l _l _l _l _l _l _l _i _l _l _i _l _l _l _l _l _l _l _l _l _l _i _i _l
(6 bi t s) (10 bi t s) (16 bi t s)
A sixteen-bit field is required to represent an integer or an atom using the
format of Section 3.1. Since a floating point number cannot be represented in
sixteen bits, in nodes that test floating point numbers, this field would hold not
the number, but the address of the number.
30 C.L. FORGY
(P Plus0x
(Goal 1' Type Simplify
(Expression ]' Name (N)
__~ . , .)
(P Time0x
(Goal 1" Type Simplify
(Expression ~" Name (N)
T Object (N))
~'Argl 0 q'Op+
1' Object (N))
q'Argl 0 l'Op*
1' Arg2 (X))
~' Arg2 (X))
ROOT FORK L003
TEQA 0, Goal
TEQA 1, Simplify
FORK L002
LO01 AND (2) = (1)
TERM Plus0x
L002 AND (2) = (1)
TERM Time0x
L003 TEQA 0, Expression
TEQN 2, 0
Fork L004
TEQA 3, +
MERGE L001
L004 TEQA 3, *
MERGE L002
; Root node of the network
; Is the element class Goal?
; Is the Type Simplify?
; Two-input node for Plus0x
; Report Plus0x is satisfied
; Two-input node for Time0x
; Report Time0x is satisfied
; Is the element class Expression?
; Is the Argl 0?
; Is the Op +?
; Is the Op *?
FIG. 2. A compiled network.
The terminal nodes coul d also be stored in single words. These nodes contain
two fields, the usual type field plus a longer field for the index or address of the
producti on that the node represents.
I I I
I TYPE [ PRODUCTION [
I I I
I_1_1_1_1_1_1_1_1_1_1_1_1_1_I_]_1_ I_1_1_1_1_1_ I_1_1_1_1_1_1_1_1_1
(6 bi t s) (26 bi t s)
The length of a two-input node would depend on the number of value pairs
tested by the node. Each node could have one word of basic information plus
one word for each value pair. The first word would contain a type field, a
pointer to the memory for the left input, a pointer to the memory for the right
input, and a field indicating how many tests are performed by the node. The
bits in the word could be allocated as follows.
I I I I I
] TYPE I COUNT I MEMORY POINTER I MEMORY POINTER I
I I I I I
I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I_I-I
(6 bi t s) (4 bi t s) (11 bi t s) (11 bi t s)
THE RETE MATCH ALGORITHM 31
The word for each test would contain three fields. Two fields would hold the
indices of the two elements to test. The remaining field would indicate the test
to perform; that is, it would indicate whether the node is to test for equality of
the two elements, for inequality, or for something else. The bits in the Word
might be allocated as follows.
I I I I
I TEST [ INDEX I INDEX [
I I I I
I_l_I_I_I_l_I_l_I_I_l_l_I_I_l_I_I_I_I_l_I_I_l_l_l_I_I_I_I_I_I_I_I
(6 bi ts) (13 bi t s) (13 bi t s)
Note that the index fields here are longer than the index fields in the one-input
nodes. This is necessary because the indices in the two-input nodes must
designate elements in the tokens as well as values in the elements.
3.3. The tokens
This section describes a space-efficient representation for tokens. This
representation is not suitable for all interpreters; it requires the interpreter to
process only one working memory change at a time, and it requires that certain
parts of the network be traversed depth first. Fortunately, these are not serious
restrictions. The simplest way to perform the match is process one token at a
time, traversing the entire network depth first. Section 4 describes an inter-
preter that operates in this manner.
If the interpreter operates this way, then it can use a stack to represent its
tokens. When a token has to be built, first the tag for the token is pushed onto
the stack, and then the working memory elements are pushed ont o the stack in
order. When tokens have to be extended (a very common operat i on--see the
code in Section 4) the additional working memory elements are just pushed
onto the stack.
The one-input nodes will be more efficient if they do not use this stack. Since
all the one-input nodes will process the same working memory el ement --t he
element that was just added to or deleted from working memor y- - t he element
should be made easily available. The element could be copied into a dedicated
location in memory, or the address of the element could be loaded into a
dedicated base register. Either of these would make it possible for the
one-input nodes to access the element without going through the stack.
3.4. The i nterpreter's state
In addition to the stack for tokens, the interpreter must maintain another stack
for its state information. One reason for the stack is to allow the interpreter to
find its way about in the network. When the interpreter passes a FORK, it
pushes the pointer it does not follow onto the stack. Then when it reaches the
32 C.L. FORGY
end of a path it pops a pointer from the stack and follows it. Anot her reason
for the stack is to provide a place for the two-input nodes to keep their local
information. As will be seen in the next section, the two-input nodes sometimes
have to suspend themselves while their successors are processed. The stack
holds the information that is needed to resume processing the two-input nodes.
4. The Network Interpreter
This section provides a concrete description of the operations performed by the
network interpreter. One node from each class has been selected, and the code
to interpret the nodes has been written. It might be noted that since the code
sequences are short and simple, they could easily be written in microcode.
The code is written in a PASCAL-like language which has literal labels and
field extraction. Field extraction is indicated by putting two numbers within
angle brackets; the first number is the index of the high order bit in the field,
and the second number is the index of the low order bit. The assumption will
be made that the bits are numbered from right to left, with the low order bit
being bit zero. Thus the expression SELF(31:26) indicates that the high order
six bits of the value of the variable SELF are to be extracted and right justified.
The main loop of the interpreter is very simple: the interpreter fetches the
next node from memory and dispatches on its type field. Let the segment of
memory that holds the nodes be called NODE_MEMORY and let the pointer
to the current node be called NC. The main loop is then:
MAIN:
SELF :--- NODE_MEMORY[ NC];
CASE SELF(31 : 26) OF !Type field is high order 6 bits
0: GOTO FORK;
1: GOTO MERGE;
2: GOTO TERM;
3: GOTO TEQA;
END;
The node is copied into the variable SELF so that the node programs can
examine it. The assignment of numbers to the various node types is arbitrary.
Goto's are used instead of procedure calls because these examples make all the
state of the interpreter explicit, and not hidden in PASCAL'S stack.
TEQA is typical of the one-input nodes for testing constants. If the segment
of memory that holds the working memory element being processed is called
CURRENT, then TEQA is as follows.
TEQA: TEMP :-- CURRENT[SELF(25:16)]: !Get the word pointed to
!by the index field
THE RETE MATCH ALGORITHM 33
IF (TEMP(31 : 16) = 0) AND
(TEMP(15:0) = SELF(15:0))
THEN GOTO SUCC
ELSE GOTO FAIL:
!Test type bits
!Test value
Either SUCC or FAI L is executed after each one-input node. SUCC is
executed when the test succeeds, and FAI L is executed when the test fails.
SUCC increments the node counter to point to the next node.
SUCC: NC:= NC+ 1;
GOTO MAIN;
FAI L tries to get a node from the stack of unprocessed nodes; if it cannot, it
halts the match. Assuming the stack is named NS and the pointer to the top of
the stack is called NSTOP, the code is:
FAIL: IF NSTOP < 0 THEN GOTO EXI T_MATCH;
NC := NS[NSTOP];
NSTOP := NSTOP- 1;
GOTO MAIN;
The one-input nodes for comparing pairs of values are similar to the other
one-input nodes. TEQS is typical of these nodes.
TEQS: IF CURRENT[SELF(25 : 16)] = CURRENT[SELF(9:0)]
THEN GOTO SUCC
ELSE GOTO FAIL;
address onto NS and then passes control to the following
FORK pushes an
node.
FORK: NSTOP : = NSTOP + 1;
NS[NSTOP] := SELF(25:0);
GOTO SUCC;
A two-input node must be able to determine whether it was reached over its
left input or its right input. This can be indicated to the node by a global
variable which usually has the value LEFT, but which is temporarily set to
RI GHT when a MERGE passes control to a two-input node. If this global
variable is called DI RECTI ON, the code for the MERGE is
MERGE: DI RECTI ON := RI GHT;
NC := SELF(25:0);
GOTO MAIN;
The two kinds of two-input nodes are very similar, so only AND is shown
here. In order not to obscure the more important information, some details of
34 C.L. FORGY
the program are omitted. The code does not show how the variables are tested,
nor does it show how tokens are added to and removed from the node's
memories. Assuming the token stack is called TS and the pointer to the top
element is called TSTOP, the program is as follows.
i Control can reach this point many times during the processing
! of a token. The node needs to update its state and put
! information on NS only once, however.
AND:IF NS[NSTOP] <> NC
THEN
BEGI N
NSTOP :-- NSTOP + 4;
NS[NSTOP] := NC;
NS[ NSTOP- 1] :-- DI RECTI ON;
NS[ NSTOP- 2] := MEMORYCONTENTS
(OPPOSITE(DIRECI'ION));
NS[ NSTOP- 3] := TSTOP;
MODI FY_MEMORY( DI RECTI ON); !Store the token
DI RECTI ON := LEFT; !Reset to the default
END;
Go process the tokens
1
LLOOP:
IF NS[NSTOP] - 1] = RI GHT THEN GOTO RLOOP
ELSE GOTO LLOOP;
Compare the token to the elements in the right memory
!If the state is not in NS
!Then put it there
Fall out of the loop when the test succeeds so that
the successors of this node can be activated
REPEAT
TEMP := NEXT_POSI TI ON( NS[ NSTOP- 2]);
IF TEMP = NIL !If right memory is empty
THEN !Then clean up and exit
BEGI N
TSTOP := NS[ NSTOP- 3];
NSTOP := NSTOP- 4;
GOTO FAI L;
END
UNTI L PERFORM AND_TEST( TEMP,LEFT);
THE RETE MATCH ALGORrl'/-IM 35
!Ext end the t oken
TSTOP := NS[ NSTOP- 3] + 1; TS[ NSTOP] := TEMP;
!Prepare NS so that cont rol will return t o this node
NSTOP := NSTOP + 1; NS[ NSTOP] := NC;
!Pass cont rol to the successors of this node
NC : = NC + SELF(25 : 22) + 1; GOTO MAI N;
! Compar e the t oken the el ement s in the left memor y
!
RLOOP:
This is similar to LLOOP.
The only remai ni ng node t ype is the TERM node. Since updat i ng the conflict
set is a l anguage-dependent operat i on, that detail of the TERM node cannot
be shown. The rest of the processing of the node is as follows.
TERM: UPDATE_CONFLI CT_SET( SELF<25 : 0));
GOTO FAI L;
5. Performance of the Algorithm
Ext ensi ve studies have been made of the efficiency of the Ret e Mat ch Al-
gori t hm. Bot h analytical studies (which det ermi ned the time and space com-
plexity of the algorithm) and empirical studies have been made. This section
present s some of the results of the analytical studies. Because of space
constraints, it was not possible to present the empirical results or the proofs of
TABLE 1. Space and time complexity
Complexity measure Best case Worst case
Effect of working memory
size on number of tokens O(1) O(W c)
Effect of production memory
size on number of nodes O(P) O(P)
Effect of production memory
size on number of tokens O(1) O(P)
Effect of working memory
size on time for one firing O(1) O(W 2c-1)
Effect of production memory
size on time for one firing O(log2 P) O(P)
C is the number of patterns in a production.
P is the number of productions in production memory.
W is the number of elements in working memory.
36 C.L. FORGY
the analytical results. The proofs and detailed results of some empirical studies
can be found in [5].
Table 1 summarizes the results of the analytical studies of the algorithm. The
usual notation for asymptotic complexity is used in this table [1].Writing that a cost
is O(f(x)) indicates that the cost varies as f(x) plus perhaps some smaller terms in
x, The smaller terms are ignored because the f(x) term will dominate when x is
large. Writing that a cost is O(1) indicates that the cost is unaffected by the factor
being considered. It should be noted that all the complexity results in Table 1
sharp; production systems achieving the bounds are described in [5].
6. Conclusions
The Ret e Match Algorithm is a met hod for comparing a set of patterns to a set
of objects in order to determine all the possible matches. It was described in
detail in this article because enough evidence has been accumulated since its
development in 1974 to make it clear that it is an efficient algorithm which has
many possible applications.
The algorithm is efficient even when it processes large sets of patterns and
objects, because it does not iterate over the sets. In this algorithm, the patterns
are compiled into a program to perform the match process. The program does
not have to iterate over the patterns because it contains a tree-structured
sorting network or index for the patterns. It does not have to iterate over the
data because it maintains state information: the program computes the matches
and partial matches for each object when it enters the data memory, and it
stores the information as long as the object remains in the memory.
Although the Ret e algorithm was developed for use in production system
interpreters, it can be used for other purposes as well. If there is anything
unusual about the pattern matching of production systems, it is only that the
pattern matching takes place on an unusually large scale. Production systems
contain rather ordinary patterns and data objects, but they contain large
numbers of them, and invocations of the pattern matcher occur very frequently
during execution. If programs of other kinds begin to use pattern matching
more heavily, they could have the same efficiency problems as production
systems, and it could be necessary to use methods like the Ret e Match
Algorithm in their interpreters as well. Certainly the algorithm should not be
used for all match problems; its use is indicated only if the following three
conditions are satisfied.
- The patterns must be compilable. It must be possible to examine them and
determine a list of features like the lists in Section 2.2.1.
- The objects must be constant. They cannot contain variables or other
non-constants as patterns can.
- The set of objects must change relatively slowly. Since the algorithm maintains
state between cycles, it is inefficient in situations where most of the data changes
on each cycle.
THE RETE MATCH ALGORITHM 37
ACKNOWLEDGMENT
The author would like to thank Allen Newell and Robert Sproull for many useful discussions
concerning this work, and Allen Newell, John McDermott, and Michael Rychener for their valuable
comments on earlier versions of this article.
REFERENCES
1. Aho, A.V., Hopcroft, J.E., and UUman, J.D., The Design and Analysis of Computer Algorithms
(Addison-Wesley, Reading, MA, 1974).
2. Cohen, B.L., A powerful and efficient structural pattern recognition system, Artificial In-
telligence 9 (1977) 223-255.
3. Forgy, C.L., A network match routine for production systems, Working Paper, 1974.
4. Forgy, C.L., A production system monitor for parallel computers, Department of Computer
Science, Carnegie-Mellon University, 1977.
5. Forgy, C.L., On the efficient implementation of production systems, Ph.D. Thesis, Carnegie-
Mellon University, 1979.
6. Forgy, C.L., o~5 user's manual, Department of Computer Science, Carnegie-Mellon
University, 1981.
7. Hayes-Roth, F. and Mostow, D.J., An automatically compilable recognition network for
structured patterns, Proc. Fourth Internat. Joint Conference on Artificial Intelligence (1975)
246-251.
8. McCracken, D., A production system version of the Hearsay-II speech understanding system,
Ph.D. Thesis, Carnegie-Mellon University, 1978.
9. McDermott, J., Neweli, A., and Moore, J., The efficiency of certain production system
implementations, in: Waterman, D.A. and Hayes-Roth, F. (Eels.), Pattern-Directed Inference
Systems (Academic Press, New York, 1978) 155-176.
10. Rychener, M.D., Production systems as a programming language for Artificial Intelligence
applications. Ph.D. Thesis, Carnegie-Mellon University, 1976.
Received May 1980; revised version received April 1981